Wood Defect Detection Report (VAE Model)

Generated on 2025-05-21 13:13:09

Model Architecture

Variational Autoencoder (VAE) with latent dimension: 64

Encoder: Custom CNN architecture with Leaky ReLU activations

Decoder: Transpose convolutions with sigmoid output activation

Anomaly score: Weighted combination of reconstruction error and KL divergence

Performance Summary

Accuracy

0.6879

F1 Score

0.6944

ROC AUC

0.6990

Optimal Threshold

0.1624

Mean IoU

0.0879

Confusion Matrix

Confusion Matrix

ROC Curve

ROC Curve

Precision-Recall Curve

Precision-Recall Curve

Anomaly Score Distribution

Score Distribution

Classification Results

Class Precision Recall F1-Score Support
Good 0.6912 0.6714 0.6812 70.0
Defect 0.6849 0.7042 0.6944 71.0

Reconstruction Examples

Image #0 - True: Good, Predicted: Defect (Score: 0.1819)

Example

Image #1 - True: Good, Predicted: Defect (Score: 0.8848)

Example

Image #2 - True: Good, Predicted: Good (Score: 0.1429)

Example

Image #70 - True: Defect, Predicted: Defect (Score: 0.1815)

Example

Image #71 - True: Defect, Predicted: Defect (Score: 0.1830)

Example

Image #72 - True: Defect, Predicted: Defect (Score: 0.1780)

Example

Segmentation Examples

Defect Image #0 - IoU Score: 0.1577

Segmentation Example

Defect Image #1 - IoU Score: 0.0302

Segmentation Example

Defect Image #2 - IoU Score: 0.0775

Segmentation Example